Research On Personalized Recommendation Based On Semi-supervised Learning

With the development of social networking, e-commerce and other Internettechnology, people gradually got into the “information overload” era from lack ofinformation era. It brings great convenience to users with vast amounts of information,but also let the user to get lost in ocean of information, it is difficult to find theinformation that they are interested in. Personalized recommended is used to pushpersonalized information for usersby mining user preferences, which is also as the mosteffective tool to solve the problem of information overload.Currently, themainstream personalized recommendation methodsinclude:collaborative filtering method and content-based method. The key of collaborativefiltering method is the similarity computing of user interest preference,which is to filterinterest items for the target user. It makes recommendationsbased on the user’s behaviorinformationmainly, but did not take advantage of the item content information and userlabel information really. Meanwhile, there aredata sparse and cold start problems.Content-basedrecommendation is an information filtering technologyin essence, whichusers only learn the history of selected items of informationsimply.It can not mining theuser feedback information on items, which often leads to excessive specialization of therecommendation result.According to the problem of above recommendation methods, the semi-supervisedlearning methods are proposedto achieve personalized recommendation based on userbehavior information and items content information. Thedetails of research work asfollows:①According to these problemsthat traditional collaborative filtering algorithm issingle in the way of calculating the similarity,the semi-supervised hybrid clusteringbased on the distance metric and Gaussian model is proposed to slove these problems.The time complexity of the traditional collaborative filtering algorithm isquadraticof thenumber of users, when the number of users is large, it is time-consuming.In this paper,the cluster analysis is used to alternative the similarity computingof user interest, whichconsiders the preferences of the user behavior and content information. Specific tocluster analysis, the algorithm takes into account not only the geometric information ofdata samples, but also take into account the normal distributioninformation of datasamples. ②According to theproblem thatthe labeled dataof user interestis too few inpersonalized recommendation, the semi-supervisedrecommended method based onactive learning and collaborative training is proposed.About thetraditionalrecommended method based onclassification model, it has very negativepotential interest problem on mining user preferences whenthe labeled datais few. In thepaper, the user behavior information and item content information is used to model,andthe unlabeled data with largest amount of information is extracted with theactivelearning strategies, which increase the sample space of training modelby query mode orlabel the unlabeled data by field experts, to improve the quality of personalizedrecommendation.③According to theproblem of it increase the burden on user or labor cost with theactive learning method, a personalized recommendation method withsemi-supervisedincremental learningbased on Gaussian symmetrical distribution isproposed.We use the large number data of no user tag information, combined with asmall amount of user tag data to build model. In the algorithm, the data selectionalgorithm chooses the unlabeled data with high confidenceand Gaussian symmetricaldistribution to iterative learning, to improve the quality of personalizedrecommendation.④According to the problem that it is difficult to measure the features vectorweights between user behavior information and item content information,thesemi-supervisedrecommended method of graph-basedis proposed, which cancalculate the weighing factors with SELF method and other methods. The algorithmconstruct the weight matrix based on nearest neighbor graph with user behaviorinformation. Specially, Sigmoid mapping function is used to measure the interest degreeof two users; wedefine the loss function of the algorithm that includes user behaviorsimilarity constraints and item content similarity constraints, and the constraints of thesetwo parts are weightedwith a balance factor.